Software Engineering Principles for Fairer Systems: Experiments with GroupCART

Fairness can be built into AI models during development, not just fixed afterward: GroupCART demonstrates that discrimination-aware machine learning is possible by redesigning how decision trees are constructed from the ground up, rather than trying to patch bias into existing models after they're trained.

Fairness and accuracy don't have to be enemies: The experiments show that fairer models can be achieved "with minimal performance degradation," meaning organizations don't necessarily have to choose between having an accurate system and a fair one—they can have both.

Customizable fairness trade-offs put control in human hands: The method allows users to adjust the balance between predictive performance and fairness based on their specific needs and values, rather than imposing a one-size-fits-all approach to what "fair enough" means.

Transparency and reproducibility matter for responsible AI: By making all code and datasets publicly available, the researchers enable independent verification and broader adoption of fairness techniques, supporting the principle that responsible AI development should be open to scrutiny.

Protected characteristics should be actively considered during model design: Rather than ignoring sensitive attributes like gender or ethnicity, GroupCART intentionally accounts for them during model construction to prevent the algorithm from inadvertently learning to discriminate against these groups.